Cervical Cancer Detection: An Enhanced Approach through Transfer Learning and Data Augmentation Using DenseNet169

© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 119
Cervical Cancer Detection: An Enhanced Approach through Transfer
Learning and Data Augmentation Using DenseNet169
Hithaishi Reddy1
1B.E., Department of Computer Science, Bangalore Institute of Technology, Bengaluru, Karnataka, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - This study presents a groundbreaking
approach to cervical cancer detection using deep learning
models, focusing on extensive datasets from Kaggle. Cervical
cancer, a significant global health concern, demands early
detection for effective treatment. Leveraging the power of
deep learning and convolutional neural networks (CNNs), I
achieved remarkable results with an accuracy of 95.27% in
cervical cancer detection. This approach involved extensive
data preprocessing and fine-tuning of a pre-trained
DenseNet architecture, resulting in an accurate and robust
model. Notably, I previously experimented with
MobileNetV2 and achieved an accuracy of 92.16%. I
conducted a comprehensive evaluation of DenseNet's
performance, revealing average precision, recall, and F1-
score values of 95.63%, 95.56%, and 95.57%, respectively,
for each of the five cervical cancer classes: Dyskeratotic,
Koilocytotic, Metaplastic, Parabasal, and Superficial-
Intermediate. The results showcase the model's impressive
capabilities across the five classes. These values
demonstrate the model's ability to correctly classify various
cervical cell types with high accuracy. This study
demonstrates the potential of deep learning models in
automating cervical cancer screening, reducing healthcare
disparities, and improving early detection rates. This
research contributes to the ongoing efforts to enhance
women's health by providing an effective and reliable tool
for cervical cancer diagnosis.
Key Words: DenseNet, CNN, Dataset,
Preprocessing, Accuracy, Precision, Recall, F1
1. INTRODUCTION
Cervical cancer remains a global health challenge,
impacting the lives of countless women across the world.
The key to effective treatment and improved patient
outcomes lies in the timely and accurate diagnosis of this
malignancy. In the current landscape of medical image
analysis, the advent of deep learning techniques has
revolutionized cancer diagnosis. In this paper, we present
a comprehensive study that focuses on the application of
transfer learning, an advanced method within the realm of
deep learning, for the detection of cervical cancer.
The SipakMed dataset, which is employed in this study, is
one of the largest and most comprehensive collections of
cervical cell images, encompassing five distinct classes:
Dyskeratotic, Koilocytotic, Metaplastic, Parabasal, and
Superficial-Intermediate. This diverse dataset allows us to
investigate the effectiveness of the DenseNet architecture
in classifying a wide range of cervical cell types. Transfer
learning is a sophisticated approach that harnesses the
knowledge acquired by a pre-trained model on a vast and
diverse dataset, such as ImageNet, and subsequently
adapts this acquired expertise to a specific and more
specialized task. This approach has gained considerable
attention within the medical field, primarily due to its
remarkable ability to achieve high accuracy in medical
image analysis, even when dealing with limited annotated
medical images.
This study is grounded in this substantial dataset, which
serves as a rich and diverse representation of cervical cell
images, enabling us to explore the integration of transfer
learning, data augmentation techniques, and specialized
model architecture, specifically the DenseNet169, to
construct a robust and accurate cervical cancer detection
system.
The key components of our investigation entail an
extensive analysis of the dataset, encompassing the
distribution of images among different classes and a
profound understanding of the variability in image
dimensions. This crucial first step is fundamental to the
design of an effective deep-learning solution. Furthermore,
we employ an array of data augmentation techniques,
including rotations, scaling, brightness adjustments, and
color manipulations, to enhance the model's capacity to
generalize, thereby reducing the risk of overfitting.
Our exploration also extends to the utilization of transfer
learning and fine-tuning. We leverage the DenseNet169
pre-trained model, which has initially been trained on
ImageNet, an extensive repository of diverse images. By
fine-tuning this pre-trained model, we harness its
acquired knowledge in recognizing intricate patterns and
adapting it to the task of cervical cancer detection.
Additionally, we address the model's training process,
which is conducted using categorical cross-entropy loss
and the Adam optimizer. To optimize training efficiency
and effectiveness, we employ several advanced techniques
such as early stopping, model checkpointing, and learning
rate reduction. In the final phase, we rigorously evaluate
the model's performance on a balanced test dataset,
visualizing the results through a confusion matrix and
detailed classification reports.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 11 | Nov 2023 www.irjet.net p-ISSN: 2395-0072

Through this rigorous and multifaceted investigation, we
achieved an exceptional accuracy rate of 95.27%,
underscoring the remarkable effectiveness of the transfer
learning approach in the realm of medical image analysis.
Our findings not only contribute to the field of cervical
cancer diagnosis but also highlight the vast potential for
transfer learning to redefine cancer detection and
transform other critical healthcare applications. In the
subsequent sections of this paper, we delve deeper into
the methodology, experimental results, and the broader
implications of our research, offering valuable insights
into the transformative power of transfer learning in the
medical domain.
2. RELATED WORKS
A few studies collectively demonstrate the potential of
machine learning and deep learning techniques in
automating and improving the accuracy of cervical cancer
detection and diagnosis, which is crucial for early
intervention and reducing the burden of this disease in
women. Paper[1] introduces a cloud-based whole-slide
imaging platform with a deep-learning classifier for dual-
stained cervical cell slides, demonstrating its ability to
reduce the number of unnecessary colposcopies and
provide consistent quality in screening. Paper[2] proposes
methods to combine features from different image
contrasts to improve the classification of cervical cancers,
showcasing the potential for automated, expert-level
diagnosis at the point of care. Paper[3] focuses on data
cleaning for AI-based diagnosis of diseases, including
cervical cancer. They present a few-shot learning-based
model that can automatically clean noisy medical image
datasets. Paper[4] explores the use of stack generalization
in classifying cervical tissue pathological images, achieving
improved classification accuracy. Paper[5] investigates the
use of deep learning models to distinguish abnormal
cervical cells from normal ones, demonstrating the
potential for automated classification with high accuracy.
Paper[6] introduces "CervDetect," which combines
machine learning algorithms to assess the risk elements of
malignant cervical formation, achieving high accuracy and
outperforming existing studies. Paper[7] presents a
method for cervical cell segmentation and classification
using deep learning models, which yields high sensitivity
and specificity for cancer detection. Paper[8] explores the
use of machine learning and deep learning algorithms for
predicting cervical cancer, with random forest, decision
tree, and other methods achieving high classification
scores. Paper[9] proposes a system that effectively
handles missing values and class imbalance in cervical
cancer datasets, achieving high accuracy with an ensemble
voting classifier.
3. METHODOLOGY
Fig -1: Proposed Architecture
The methodology for cervical cancer detection using
DenseNet encompasses a systematic approach to
harnessing deep learning for accurate diagnosis. The data
obtained is meticulously organized into distinct cell
classes. Preprocessing techniques standardize image
attributes and facilitate uniformity. Data is then
thoughtfully split into training, validation, and test sets to
ensure model training and evaluation integrity.
Augmentation diversifies the training data. Leveraging the
DenseNet architecture, transfer learning is employed with
pre-trained weights fine-tuned on the cervical cell dataset.
Extensive model training, with an emphasis on
hyperparameter tuning, is conducted. The model's
performance is rigorously evaluated using a test dataset.
The outcome is analyzed through a confusion matrix and a
detailed classification report, offering insights into its
ability to accurately classify cervical cell types. The results
highlight the model's precision and potential clinical
utility, paving the way for early intervention in cervical
cancer diagnosis.
3.1 DATA COLLECTION
The dataset was obtained from the "Cervical Cancer
largest dataset (SipakMed)" available on Kaggle. The data
was organized into distinct classes, each representing
different types of cervical cells, namely Dyskeratotic,
Koilocytotic, Metaplastic, Parabasal, and Superficial-
Intermediate. Additionally, data preprocessing techniques
were applied to standardize image sizes and enhance
image quality. The dataset was further divided into
training, validation, and test sets, ensuring that the model
would be trained and evaluated on separate, non-
overlapping data subsets.

3.2 IMAGE PREPROCESSING
Multiple image preprocessing techniques were
meticulously applied to enhance the quality and suitability
of the input data, a crucial step in enabling the neural
network to effectively learn from cervical cell images.
Initially, the images underwent uniform resizing to a
standard dimension of 110x110 pixels, ensuring
consistency in their sizes. Subsequently, an array of data
augmentation methods was employed to enrich the
training dataset. These techniques encompassed
operations such as rotation, scaling, translation, shearing,
and flipping (both horizontally and vertically).
Additionally, brightness and contrast adjustments, as well
as the introduction of various types of noise, contributed
to data diversification.
Furthermore, color normalization techniques were
implemented to maintain uniformity in color balance,
contrast, and brightness across all images. The utilization
of histogram equalization played a significant role in
enhancing image contrast, thus ensuring that subtle
variations in the images were adequately represented. As
a final step, the images were transformed into NumPy
arrays, facilitating seamless integration with machine
learning frameworks. Class labels underwent one-hot
encoding, representing each class as a binary vector.
Collectively, these preprocessing measures impeccably
readied the data for the DenseNet model, effectively
addressing issues related to data quality, variability, and
format. The comprehensive preprocessing pipeline
significantly bolstered the model's performance in
accurately classifying diverse cervical cell types.
Fig -3: Sample Image Dimensions
Fig -4: Representation of Image Sizes
3.2.1. DATA AUGMENTATION
In the cervical cancer detection code utilizing DenseNet, a
comprehensive set of data augmentation techniques was
employed to augment the training dataset, thereby
enhancing the model's ability to generalize effectively and
cope with variations in the input data. These
augmentation techniques encompassed a range of
transformations, including random rotations within -90 to
+90 degrees, scaling variations, translations along both X
and Y axes, shearing with angles of -2 to +2 degrees, and
horizontal and vertical flipping. Brightness and contrast
adjustments simulated diverse lighting conditions, while
the addition of Gaussian and Laplace noise introduced
variability in pixel values, mimicking real-world noise or
image quality issues. These augmentation strategies
collectively contributed to a more extensive and diverse
training dataset, reducing the risk of overfitting and
empowering the model to better classify different types of
cervical cells accurately.
Fig -2: Number of images in each class

Fig -5: Density Plot of Images Sizes
3.3 DENSENET-169
In the implementation of DenseNet for cervical cancer
detection, the DenseNet-169 architecture was utilized.
First, a pre-trained DenseNet model with weights obtained
from ImageNet was imported. The model's architecture
included multiple convolutional layers, dense blocks,
transition layers, and a global average pooling layer. Fine-
tuning was performed to adapt the pre-trained model to
the specific task of cervical cancer classification. The
model's layers were divided into two groups: the first 249
layers were kept frozen (non-trainable), while the
subsequent layers were trainable. This approach allowed
the network to retain knowledge from ImageNet while
learning to extract features relevant to cervical cell image
classification. Customized fully connected layers were
added to the model's top, including a dense layer with
1024 units and a ReLU activation function, followed by an
output layer with five units, corresponding to the five
cervical cell classes. The model was then compiled with an
Adam optimizer, categorical cross-entropy loss function,
and accuracy as the evaluation metric. During training,
callbacks, including early stopping, learning rate
reduction, and model checkpointing, were applied to
monitor and improve model performance. The model was
subsequently trained on the prepared cervical cell image
dataset, with notable success, achieving an impressive
accuracy of 95.27% in accurately classifying cervical cell
types, making it a powerful tool for cervical cancer
detection.
Fig -6: Confusion Matrix
4. RESULTS AND ANALYSIS
The code ensures a balanced dataset splitting and
conducts a rigorous evaluation. Achieving an accuracy of
95.27%, the model demonstrates its effectiveness in
recognizing cervical cancer patterns in medical images, for
early cancer detection. Early stopping is a valuable
training technique that halts the model's training and
hence runs only for 8 epochs out of the planned 20 epochs.
This was because the model's validation loss hadn't
improved for three consecutive epochs, and further
training wasn't likely to enhance performance. Early
stopping is efficient, saving both time and computational
resources, and it prevents overfitting.
Fig -7: Performance Analysis
5. CONCLUSION
In conclusion, this study demonstrates the effectiveness of
DenseNet in cervical cancer detection, showcasing a high
accuracy of 95.27% on the SipakMed dataset. The
implementation of comprehensive data preprocessing
techniques, including image resizing, data augmentation,
color normalization, and label encoding, ensures the
model's readiness to handle diverse cervical cell images.

The use of transfer learning, fine-tuning a pre-trained
DenseNet, further enhances the model's performance,
suggesting its practical applicability in clinical settings. By
automating the classification of cervical cell images, this
research contributes to improving the accuracy, efficiency,
and accessibility of cervical cancer diagnosis, offering a
promising solution for the early detection of this prevalent
and life-threatening disease.
6. FUTURE WORKS
Future work in cervical cancer detection using DenseNet
should focus on diversifying and expanding datasets for
improved generalization. Developing user-friendly
software for automated screening is essential. Addressing
class imbalance through SMOTE and enhancing real-time
image capture and telemedicine integration can extend
access to underserved regions. Improving model
interpretability is crucial for clinical adoption.
Additionally, the model's adaptability to tasks beyond
detection, such as disease staging, offers promising
research opportunities.
REFERENCES
[1] Wentzensen, Nicolas, et al. "Accuracy and efficiency of
deep-learning–based automation of dual stain cytology in
cervical cancer screening." JNCI: Journal of the National
Cancer Institute 113.1 (2021): 72-79.
[2] Asiedu, Mercy N., et al. "Combining multiple contrasts
for improving machine learning-based classification of
cervical cancers with a low-cost point-of-care Pocket
colposcope." 2020 42nd Annual International Conference of
the IEEE Engineering in Medicine & Biology Society (EMBC).
IEEE, 2020.
[3] Bijoy, M. B., et al. "Deep Cleaner—A Few Shot Image
Dataset Cleaner Using Supervised Contrastive
Learning." IEEE Access 11 (2023): 18727-18738.
[4] Zhang, Shuailei, et al. "Research on application of
classification model based on stack generalization in the
staging of cervical tissue pathological images." IEEE
Access 9 (2021): 48980-48991.
[5] Yu, Suxiang, et al. "Automatic classification of cervical
cells using deep learning method." IEEE Access 9 (2021):
32559-32568.
[6] Mehmood, Mavra, et al. "Machine learning assisted
cervical cancer detection." Frontiers in Public Health 9
(2021): 788376.
[7] Allehaibi, Khalid Hamed S., et al. "Segmentation and
classification of cervical cells using deep learning." IEEE
Access 7 (2019): 116925-116941.
[8] Al Mudawi, Naif, and Abdulwahab Alazeb. "A model for
predicting cervical cancer using machine learning
algorithms." Sensors 22.11 (2022): 4132.
[9] Karamti, Hanen, et al. "Improving Prediction of
Cervical Cancer Using KNN Imputed SMOTE Features and
Multi-Model Ensemble Learning Approach." Cancers 15.17
(2023): 4412.
BIOGRAPHY
Hithaishi Reddy is currently a
student in Bangalore Institute of
Technology, Bengaluru,
Karnataka, India, pursuing her
Bachelor’s in Computer Science
and Engineering.

Cervical Cancer Detection: An Enhanced Approach through Transfer Learning and Data Augmentation Using DenseNet169

More Related Content

Similar to Cervical Cancer Detection: An Enhanced Approach through Transfer Learning and Data Augmentation Using DenseNet169 (20)

More from IRJET Journal (20)

Recently uploaded (20)

Cervical Cancer Detection: An Enhanced Approach through Transfer Learning and Data Augmentation Using DenseNet169